This manuscript describes the development of neural networks to replace the aerosol optics in a climate model with a more detailed treatment, which is based on running the same types of codes with more detail to develop a parameterization which better represents more detailed modeling. In general the manuscript is well written and clear. There is a nice discussion of how the neural network is developed. However,  my main critique is that the evaluation section (section 5) is pretty minimal. Just some error curves. What does it look like in the full model? You have demonstrated that the new parameterization represents the more detailed code better than the existing parameterization. Does it change the answers in the climate model it is designed for in any meaningful way, and does it cost anything more to run it. Also good to note in the conclusions what lessons you learn from this experience about building neural networks for parameterization replacement. This is probably suitable for publication with minor revisions, but with at least trying it in a climate model perhaps. 

Detailed comments below.

Page 1, L10: Would be good to have more detail on what ‘outperform’ means specifically in another sentence or two.

Page 1, L15: Disingenuous. The direct effects of aerosols are not the largest uncertainty: only indirect effects on clouds. 

Page 2, L29: Example of climate models generating training data (for replacing part of a parameterization: Gettelman et al 2021.

Gettelman, A., D. J. Gagne, C.-C. Chen, M. W. Christensen, Z. J. Lebo, H. Morrison, and G. Gantos. “Machine Learning the Warm Rain Process.” Journal of Advances in Modeling Earth Systems 13, no. 2 (2021): e2020MS002268. https://doi.org/10.1029/2020MS002268.

Page 3, L63: clarify ‘these optical properties (absorption…etc’

Page 3, L77: what is the size range here? Please be explicit.

Page 3, L90: The CESM reference should probably be Danabasoglu, et al 2020.

Danabasoglu, G., J.-F. Lamarque, J. Bacmeister, D. A. Bailey, A. K. DuVivier, J. Edwards, L. K. Emmons, et al. “The Community Earth System Model Version 2 (CESM2).” Journal of Advances in Modeling Earth Systems 12, no. 2 (2020): e2019MS001916. https://doi.org/10.1029/2019MS001916.

Page 5, L148: how much error is there in the approximations? Can you quantify it?

Page 9, L251: why would a random network do better? Is there an explanation? Isn’t that a form of overfitting?

Page 10, L289: Please describe these terms a bit. ReLU, ELU, Leaky ReLU and Parametric ReLU

Page 10, L291: what is a transfer function? Above you call them activation functions. Please clarify.

Page 11, Figure 2: is the red dot the ‘optimum’ network?

Page 12, L350: can you make the evaluation a bit more quantitative in spots? It seems a bit ‘weak’ right now, especially compared to the rest of the paper. 

Page 13, L360: What does the first column (Table:) of table 1 mean? Should it say something?

Page 13, L364: Can you explain the patters in Figure 5? What do they arise from?

Page 16, L402: Other lessons learned? It would be great to share in the paper. 

Review of the manuscript egusphere-2022-559 by Geiss et al.

The manuscript (MS) presents a new method (based on Artificial Neural Networks (ANNs)) for online calculation of the optical properties of the internally mixed aerosols. Current parametrizations and look-up tables are either computationally unaffordable or fail to capture the large variabilities in aerosol properties. The training dataset is based on the Mie code that directly computes the optical properties of aerosols by considering the variability of the particle sizes, wavelengths, and refractive indices. This approach is similar to previous parameterizations but uses a higher resolution for different parameters. By evaluating ANNs with randomly generated wirings, the optimal network architectures are identified for SW and LW. The results show that randomly generated deep ANNs lead to lower error compared to the conventional multi-layer perceptron. Besides, the ANN-based parameterization outperforms the current parameterization.

The paper is very well structured and written. I really enjoyed the detailed explanations of the assumptions and methods that makes it easy to follow the results. The methods and results are robust with major benefits for the aerosol modeling community. Thus, I recommend publication after addressing the minor points/questions listed below.

1. With respect to the I/O, it is not clear why nine variables are chosen. Any pre-processing or input selection procedure? Especially two parameters “surface mode radius over wavelength” and “surface mode radius” are obviously correlated. This should not happen.
2. Why do you need one-hot encoding? What additional information does it contain for the modes? 
3. It can be expected that 2-3 hidden layers can capture the nonlinearities of the system very well and more hidden layers often lead to over-fitting (shown in Fig 2). But it is not clear if the rather minor MAE reduction by random wired networks (outperform is too strong here) is justified by its computational costs/complexity.
4. I would like to see ANN vs. Mie similar to figure 5 but for all parameters: extinction coefficient, single scattering albedo and asymmetry parameter in SW and WL.